Install Extras Matrix

This page lists every optional extra — for each, the component slot it serves, the implementation(s) it enables, and the packages it pulls in.

The guiding rule: pip install indx stays light. Core carries no cloud SDKs, no vector-DB clients, and no heavy parsing toolchains. Anything expensive is opt-in. The documented model defaults use cloud adapters once their extras are installed, and the local profile stays available for air-gapped runs.

Related references

• Installation covers install paths.
• Registry & Defaults covers how a name like parser = "docling" resolves to an implementation.

What the bare install pulls

A plain pip install indx carries only a small set of pure-Python (and one compiled-wheel) dependencies:

Package	Why it’s in core
Pydantic v2	All boundary-crossing data models (config, manifest, documents, chunks, relations)
Typer	Builds the typed CLI command tree
Rich	Terminal rendering — progress bars, tables for indx inspect, tracebacks
Click	Underlies Typer’s parsing and completion
pydantic-settings	Layers config sources: defaults → indx.toml → env vars → CLI flags

TOML parsing needs no dependency. tomllib has been in the Python standard library since 3.11, and indx requires Python 3.11–3.13.

pydantic-core is a wheel

Pydantic v2 brings in the compiled pydantic-core. It is small and has wheels for every supported platform. Every serialized artifact depends on it for correctness, so it stays in core rather than behind an extra.

Zero-dependency fallbacks (ship in core)

The bare install is not just a skeleton. It ships working fallbacks for every slot that would otherwise need an extra. A full end-to-end run is therefore possible offline with nothing but pip install indx:

Slot	Fallback (in core)	Behaviour
parser	plaintext	Reads files as plain text — no Docling/Torch needed
store	jsonl	Vectors and metadata written inline; brute-force linear search
vlm	none	Skips vision enrichment entirely (this is also the default VLM)
output writer	.indx and jsonl	Self-contained archive, or newline-delimited documents/chunks

Air-gapped by design

With --offline (which fills the in-core plaintext parser, none LLM/VLM, hash embedder, and jsonl store), indx <dir> --out <dir> --offline produces a usable knowledge space with no database installed and no network access. A bare indx <dir> --out <dir> uses the default cloud stack (Docling, OpenAI, Qdrant) and needs the relevant extras plus an API key — running it offline without those flags fails by design and prints a hint pointing you to --offline. For richer parsing, real ANN search, or LLM enrichment, add the relevant extras below. See the Local & air-gapped guide.

The full extras matrix

pyproject.toml maps each extra to the packages its implementation needs, under [project.optional-dependencies]. The “Pulls” column shows the headline dependency each extra brings in.

pip install …	Slot	Implementation(s) enabled	Pulls (illustrative)
indx[docling]	parser	DoclingParser	docling
indx[unstructured]	parser	UnstructuredParser	unstructured
indx[llamaparse]	parser	LlamaParseParser	llama-cloud-services
indx[markitdown]	parser	MarkItDownParser	markitdown
indx[ollama]	llm / vlm	OllamaLLM	ollama
indx[vllm]	llm	VLLMClient	vllm, openai (OpenAI-compatible API)
indx[openai]	llm / vlm / embed	OpenAILLM, GPT4oVLM, OpenAIEmbedder	openai
indx[anthropic]	llm	AnthropicLLM	anthropic
indx[litellm]	llm / embed	LiteLLMClient, LiteLLMEmbedder	litellm (one client for 100+ providers, on-prem + AWS/Azure/GCP/cloud)
indx[azure]	llm / vlm / embed / parser / store	AzureOpenAILLM, AzureOpenAIVLM, AzureOpenAIEmbedder, DocumentIntelligenceParser, AzureAISearchStore	openai, azure-ai-documentintelligence, azure-search-documents, azure-identity, azure-core
indx[qwen-vl]	vlm	QwenVLClient	transformers, torch, qwen-vl-utils
indx[vlm-local]	vlm	LocalVLM	httpx
indx[bge]	embed	BGEM3Embedder (name: bge-m3)	FlagEmbedding, torch
indx[e5]	embed	E5Embedder	sentence-transformers, torch
indx[cohere]	embed	CohereEmbedder	cohere
indx[qdrant]	store	QdrantStore	qdrant-client
indx[pgvector]	store	PgVectorStore	psycopg[binary], pgvector
indx[chroma]	store	ChromaStore	chromadb
indx[lancedb]	store	LanceDBStore	lancedb
indx[aws]	parser / llm / vlm / embed / store	TextractParser, BedrockLLM, BedrockVLM, BedrockEmbedder, S3VectorsStore	boto3>=1.40
indx[aws-opensearch]	store	OpenSearchStore	indx[aws], opensearch-py>=2.4
indx[gcp]	parser / llm / vlm / embed / store	DocumentAIParser, VertexLLM, VertexVLM, VertexEmbedder, BigQueryStore	google-genai>=1.0, google-cloud-documentai>=2.20, google-cloud-bigquery>=3.20
indx[gcp-vectorsearch]	store	VertexVectorStore	indx[gcp], google-cloud-aiplatform>=1.50
indx[langchain]	output / agent	LangChainWriter; LangChain agent tools + retriever	langchain-core
indx[llamaindex]	output	LlamaIndexWriter	llama-index-core
indx[mcp]	agent	MCP server (indx mcp); KnowledgeConnector.mcp()	fastmcp>=2.0, falls back to the FastMCP bundled in the mcp SDK
indx[pydantic-ai]	agent	Pydantic AI tools	pydantic-ai-slim
indx[openai-agents]	agent	OpenAI Agents SDK function tools	openai-agents
indx[claude-agent]	agent	Claude Agent SDK in-process MCP server	claude-agent-sdk
indx[agent]	agent (bundle)	every framework adapter + the MCP server	union of the agent extras + langchain
indx[local] / indx[defaults]	bundle	docling + ollama + bge-m3 + qdrant (the local profile)	the local / air-gapped stack
indx[all]	bundle	everything above + all cloud stacks	union of all extras (including aws, aws-opensearch, azure, gcp, gcp-vectorsearch)

Shared extras

A single extra can serve more than one slot when the same SDK is reused. indx[openai] enables the OpenAI LLM, the GPT-4o VLM, and the OpenAI embedder, all via the openai package. indx[ollama] covers both LLM and, where applicable, VLM use through the local Ollama runtime.

Bundles

Two convenience bundles install groups of extras in one command:

Bundle	Installs	Use when
indx[local]	docling + ollama + bge + qdrant	You want the local profile in one line.
indx[all]	The union of every extra above	CI, exploration, or trying every backend.

indx[local] and indx[defaults] are aliases for the same local bundle. Installing either pulls the local profile: Docling, Ollama qwen2.5, BGE-M3, and Qdrant.

# Opt-in local profile (air-gapped stack): Docling, Ollama (qwen2.5), BGE-M3, Qdrant
# indx[local] and indx[defaults] are aliases for the same bundle
pip install "indx[local]"

# Everything, for CI or experimentation
pip install "indx[all]"

# Or compose exactly what you need
pip install "indx[markitdown,openai,chroma]"

Quote the bracket

Some shells (notably zsh) treat [ and ] as glob characters. Quote the spec — pip install "indx[local]" — to avoid a “no matches found” error.

AI-agent connectors

The indx.agent package turns a built .indx archive into live tools for an AI agent. Each framework adapter is gated behind its own extra; indx[agent] installs them all (plus the MCP server).

pip install "indx[agent]"           # every adapter + the MCP server
pip install "indx[mcp]"             # just the universal MCP server (Mastra, Cursor, Claude Desktop…)

The langchain extra does double duty: it powers both the build-time langchain output writer and the LangChain agent connector (both ride on langchain-core). See the AI agents guide.

Cloud profile extras

Three extras bundle a complete single-vendor pipeline — parser, LLM, VLM, embedder, and vector store — all satisfied by one cloud’s managed services. Install one extra, pass the matching CLI flag, and every slot is filled automatically (explicit flags still win; see CLI reference).

pip install "indx[aws]"    # Textract · Bedrock · Titan · S3 Vectors
pip install "indx[azure]"  # Document Intelligence · Azure OpenAI · Azure AI Search
pip install "indx[gcp]"    # Document AI · Vertex AI (Gemini) · gemini-embedding-001 · BigQuery

SDK version pins

Each cloud’s SDK is gated behind an extra and is never pulled by a bare pip install indx.

Extra	Key SDK dependency	Notes
indx[aws]	boto3>=1.40	boto3 is the single SDK for Bedrock, Textract, and S3 Vectors. boto3>=1.40 is required for the S3 Vectors client (GA 2025-12). Older versions lack the s3vectors service client and raise a clear upgrade prompt at runtime.
indx[aws-opensearch]	opensearch-py>=2.4	Adds the opt-in OpenSearch Serverless store on top of indx[aws]. Use this when you need lower-latency ANN search.
indx[azure]	openai>=1.40, azure-ai-documentintelligence>=1.0.0, azure-search-documents>=11.5.0, azure-identity>=1.16, azure-core>=1.30	Widened from the previous ["openai"] definition. The extra now covers the full Azure stack. The openai package stays at the same pin it already required; the Azure service SDKs are additive. azure-ai-documentintelligence>=1.0.0 is required for the body= argument shape (pre-1.0.0 betas used a different parameter). azure-search-documents>=11.5.0 is required for the vector-search profile classes. azure-identity enables DefaultAzureCredential (managed-identity / az login) when no API-key env var is set.
indx[gcp]	google-genai>=1.0, google-cloud-documentai>=2.20, google-cloud-bigquery>=3.20	google-genai is the unified SDK for Vertex AI Gemini (LLM, VLM, and embedder slots). The deprecated vertexai.generative_models classes are not used.
indx[gcp-vectorsearch]	google-cloud-aiplatform>=1.50	Adds Vertex Vector Search on top of indx[gcp]. Vertex Vector Search requires deploying an index to an endpoint before querying (a 20–60 minute operation); the default GCP store (bigquery) has no such requirement.

Default and opt-in stores per cloud

Cloud	Default store	Opt-in store	Why the split
AWS	s3vectors (indx[aws])	opensearch (indx[aws-opensearch])	S3 Vectors needs only boto3; OpenSearch Serverless requires opensearch-py plus extra IAM data-access policies. S3 Vectors is zero-infra and ~90 % cheaper; OpenSearch gives lower latency and higher QPS.
Azure	azure-search (indx[azure])	— (Azure AI Search is the only native option)	All Azure service SDKs are already in indx[azure].
GCP	bigquery (indx[gcp])	vertex-vector (indx[gcp-vectorsearch])	BigQuery is serverless with no endpoint to deploy; Vertex Vector Search requires a long-running, always-billed endpoint. BigQuery maps cleanly to upsert/search/delete at seconds-scale latency.

Credentials stay in the environment

Cloud adapters never read or log API keys directly. Each cloud’s native credential chain is the source of truth: the boto3 chain (AWS_* env / profile / role) for AWS; an API key env var or DefaultAzureCredential for Azure; Application Default Credentials (GOOGLE_APPLICATION_CREDENTIALS / attached SA / gcloud auth) for GCP. Config sub-tables carry only non-secret knobs (region, endpoint, model id, index name).

Missing extras: MissingDependencyError

Because the registry only ever imports concrete implementations lazily, a missing extra never breaks an unrelated code path. The error surfaces only when you actually select that slot.

When an implementation module is imported without its extra installed, the utils.lazy.require_extra(...) helper raises a single, actionable error. Selecting the Qdrant store without indx[qdrant] installed yields:

MissingExtraError: store 'qdrant' requires the 'qdrant' extra: pip install indx[qdrant]

The message always names the slot, the selected implementation, and the exact pip install indx[...] command to fix it. MissingDependencyError is part of the shared exception hierarchy rooted at IndxError; see Errors & exit codes.

Errors are slot-scoped, not import-scoped

Installing a plugin or selecting a default does not eagerly import every backend. If you run with the in-core plaintext parser and jsonl store, no parser/store extra is ever touched — so an air-gapped install simply never triggers a MissingDependencyError. The error appears at the moment a slot resolves to an implementation whose extra is absent.

Picking extras per slot

If you’d rather choose deliberately than install a bundle, each slot has a dedicated guide that weighs the options:

• Parsers: Choosing a parser
• LLM / VLM enrichment: Enrichment with LLM & VLM
• Embedders: Choosing an embedder
• Stores: Choosing a store
• Output writers: Output formats

For how a name in indx.toml resolves to one of these classes — and how third-party plugins register their own — see Registry & Defaults and Configuration.